Automating Submissions

Do NOT submit segments if there is an existing segment that is good. If you are found to be automating segment submissions, you will be warned and/or banned.

You can use timings provided from automated tools but you MUST manually review & preview them.

ML/ AI Segments

While there are multiple ML/ AI Projects based around SponsorBlock (https://github.com/xenova/sponsorblock-ml and https://github.com/andrewzlee/NeuralBlock), Automating submissions is strongly discouraged. While these models are accurate to the transcriptions, they are not trained to include frame buffers, transitions or segues and often include too much.

Repeating/ Consistent Timings

Personally, I have made segments for playlists and for channels and made similarly timed segments but the consistency and accuracy of "repeating" intros doesn't account for early fade-ins, interruptions or changes in styling over the year(s). 90% accuracy for 90% of videos isn't good enough.

Using LLMs to generate DeArrow titles

While it might be tempting to download the video transcript and ask a LLM to generate a title for the video, this approach is prone to errors and has various edge cases that will cause the titles generated to be useless. And besides that, titles that LLMs output often don't follow the DeArrow guidelines.

Creating quality titles should be prioritized over submitting bad titles quickly.

Automation Hall of Shame

Below are some examples how submission automation with LLMs went wrong.

• 
  A burst of errors found in the history of submissions of an automated user
• 
  Another burst of errors
• 
  A LLM that could only see the video transcript labeled many YouTube shorts without speech as "music videos"
• 
  Another example of the same edge case
• 
  LLMs sometimes just output pure nonsense